Build up the Bayesian network

loading the package

library(ggplot2)

Need help getting started? Try the R Graphics Cookbook: https://r-graphics.org

library(gplots)

Attaching package: ‘gplots’

The following object is masked from ‘package:stats’:

    lowess

library(bnlearn)

Load the data

source('loadHNSCCdata.R')

Attaching package: ‘dplyr’

The following objects are masked from ‘package:stats’:

    filter, lag

The following objects are masked from ‘package:base’:

    intersect, setdiff, setequal, union

Loading required package: nationalparkcolors

library(reshape)

Attaching package: ‘reshape’

The following objects are masked from ‘package:tidyr’:

    expand, smiths

The following object is masked from ‘package:dplyr’:

    rename

tumProtTable <- cast(tumProts, Gene~patient, mean)

Using inferredTumAbund as value column.  Use the value argument to cast to override this choice

rownames(tumProtTable) <- paste0(tumProtTable$Gene, "_T")
tumProtTable$Gene <- NULL
normProtTable <- cast(normProts, Gene~patient, mean)

Using inferredNormAbund as value column.  Use the value argument to cast to override this choice

rownames(normProtTable) <- paste0(normProtTable$Gene, "_N")
normProtTable$Gene <- NULL
imgFeatTable <- cast(imgFeat, Feature_Type~patient, mean)
rownames(imgFeatTable) <- imgFeatTable$Feature_Type
imgFeatTable$Feature_Type <- NULL

lnTableAll <- red.dat
lnTableAll$affected <- abs(unclass(factor(lnTableAll$lnSpread)) - 2)
lnTableAll$NoLN <- abs(unclass(factor(lnTableAll$lnCounts)) - 3)
lnTable <- as.data.frame(lnTableAll[ ,c("patient", "affected", "NoLN")])
lnAfTable <- as.data.frame(lnTableAll[ ,c("patient", "affected")])
lnNoTable <- as.data.frame(lnTableAll[ ,c("patient", "NoLN")])

rownames(lnTable) <- lnTable$patient
rownames(lnAfTable) <- lnTable$patient 
rownames(lnNoTable) <- lnTable$patient
lnTable$patient <- NULL 
lnAfTable$patient <- NULL 
lnNoTable$patient <- NULL
lnTableF <- as.data.frame(t(lnTable))
lnAfTableF <- as.data.frame(t(lnAfTable))
lnNoTableF <- as.data.frame(t(lnNoTable))

load("./analysis_SF/allCorrs.RData")

features <- c("Complexity", "InterquartileRange", "LeastAxisLength", "MajorAxisLength", "Maximum2DDiameterRow", "Maximum2DDiameterSlice", "MeshVolume", "SizeZoneNonUniformity", "SmallDependenceHighGrayLevelEmphasis", "SurfaceArea", "TotalEnergy", "VoxelVolume")

imgFeatTableF <- imgFeatTable[features, ]
imgFeatTableF

load("./analysis_SF/allRawCorrs.RData")

filterCorr <- function(corTable, th = 0.6) {
  return(as.data.frame(corTable[,colSums((corTable > th) | (corTable < -th))> 0]))
}

corrIfTpF <- filterCorr(corrIfTp[features,], 0.5) # 0.5
tumSelected <- paste0(colnames(corrIfTpF), "_T")
print(length(tumSelected))

[1] 8

tumProtTableF <- tumProtTable[tumSelected,]
corrIfNpF <- filterCorr(corrIfNp[features,], 0.6) # 0.6
normSelected <- paste0(colnames(corrIfNpF), "_N")
print(length(normSelected))

[1] 18

normProtTableF <- normProtTable[normSelected,]

tumProtTableF

normProtTableF

col_names <- Reduce(intersect, list(colnames(imgFeatTableF), colnames(tumProtTableF), colnames(normProtTableF), colnames(lnTableF)))
imgFeatProtTable <- data.frame(t(data.frame(rbind(imgFeatTableF[, col_names], tumProtTableF[, col_names], normProtTableF[, col_names], lnTableF[, col_names]))))
imgFeatProtTable

rho = cor(imgFeatProtTable)
heatmap.2(rho, scale = "none", trace = "none", revC = TRUE)#, breaks = palette.breaks)

ug = empty.graph(colnames(rho))
amat(ug) = (rho > 0.6) + 0L - diag(1L, nrow(rho))
graphviz.plot(ug, layout = "fdp", shape = "ellipse")

bl <- rbind(
    set2blacklist(c(rownames(tumProtTableF), rownames(normProtTableF))), 
    set2blacklist(rownames(lnTableF)), 
    set2blacklist(rownames(imgFeatTableF)), 
    tiers2blacklist(list(c(rownames(tumProtTableF), rownames(normProtTableF)),  
                         rownames(imgFeatTableF))), 
    tiers2blacklist(list(rownames(lnTableF), rownames(imgFeatTableF))),
    tiers2blacklist(list(c(rownames(tumProtTableF), rownames(normProtTableF)), 
                         rownames(lnTableF)))#,
    # tiers2blacklist(list(c(rownames(tumProtTableF), rownames(normProtTableF)), 
    #                      rownames(imgFeatTableF)))
    )
wl <- rbind(
    tiers2blacklist(list(rownames(lnTableF), 
                         c(rownames(tumProtTableF), rownames(normProtTableF)))),
    # tiers2blacklist(list(rownames(imgFeatTableF), 
    #                      c(rownames(tumProtTableF), rownames(normProtTableF)))),
    tiers2blacklist(list(rownames(imgFeatTableF), rownames(lnTableF)))
    )

str.raw = boot.strength(imgFeatProtTable, R = 100, algorithm = "hc", algorithm.args = list(blacklist = bl, whitelist = wl))
attr(str.raw, "threshold")

[1] 0.5

avg.raw.full = averaged.network(str.raw)

strength.plot(avg.raw.full, str.raw, shape = "ellipse")#, highlight = list(arcs = wl))

strength.plot(avg.raw.full, str.raw, threshold = 0.85, shape = "ellipse")#, highlight = list(arcs = wl))

avg.raw.full$learning$whitelist = wl
avg.raw.full$learning$blacklist = bl
nrow(undirected.arcs(cpdag(avg.raw.full, wlbl = TRUE)))

[1] 0

nrow(str.raw[with(str.raw, strength > 0.50 & direction > 0.50), ])

[1] 261

nrow(str.raw[with(str.raw, strength > 0.85 & direction > 0.50), ])

[1] 78

min(str.raw[with(str.raw, strength > 0.50 & direction > 0.50), "direction"])

[1] 1

avg.raw.simpler = averaged.network(str.raw, threshold = 0.8)
g <- Rgraphviz::layoutGraph(bnlearn::as.graphNEL(avg.raw.simpler))
graph::nodeRenderInfo(g) <- list(fontsize=100)
Rgraphviz::renderGraph(g)

strength.plot(avg.raw.simpler, str.raw, shape = "ellipse")#, highlight = list(arcs = wl))

dag = hc(imgFeatProtTable, whitelist = wl, blacklist = bl)
dag

  Bayesian network learned via Score-based methods

  model:
   [DCUN1D1_T][EIF4H_T][KRAS_T][MLF2_T][PDAP1_T][PUM1_T][SUMF2_T][USP30_T][ATP8A1_N][ATP9B_N][CDC25B_N][CDK14_N][COL21A1_N][DTD2_N][ERP44_N]
   [FAM171A2_N][FAM204A_N][GORASP2_N][IAH1_N][NCEH1_N][PDE3A_N][PHLPP1_N][PLEKHA3_N][RPL9_N][SPN_N][STK39_N]
   [affected|DCUN1D1_T:EIF4H_T:KRAS_T:MLF2_T:PDAP1_T:PUM1_T:SUMF2_T:USP30_T:ATP8A1_N:ATP9B_N:CDC25B_N:CDK14_N:COL21A1_N:DTD2_N:ERP44_N:FAM171A2_N:FAM204A_N:GORASP2_N:IAH1_N:NCEH1_N:PDE3A_N:PHLPP1_N:PLEKHA3_N:RPL9_N:SPN_N:STK39_N]
   [NoLN|DCUN1D1_T:EIF4H_T:KRAS_T:MLF2_T:PDAP1_T:PUM1_T:SUMF2_T:USP30_T:ATP8A1_N:ATP9B_N:CDC25B_N:CDK14_N:COL21A1_N:DTD2_N:ERP44_N:FAM171A2_N:FAM204A_N:GORASP2_N:IAH1_N:NCEH1_N:PDE3A_N:PHLPP1_N:PLEKHA3_N:RPL9_N:SPN_N:STK39_N]
   [Complexity|KRAS_T:USP30_T:RPL9_N:affected:NoLN][InterquartileRange|KRAS_T:STK39_N:affected:NoLN]
   [LeastAxisLength|MLF2_T:USP30_T:ATP8A1_N:ATP9B_N:IAH1_N:SPN_N:affected:NoLN]
   [MajorAxisLength|DCUN1D1_T:KRAS_T:MLF2_T:SUMF2_T:USP30_T:ATP9B_N:CDC25B_N:DTD2_N:ERP44_N:NCEH1_N:SPN_N:STK39_N:affected:NoLN]
   [Maximum2DDiameterRow|USP30_T:COL21A1_N:NCEH1_N:RPL9_N:affected:NoLN]
   [Maximum2DDiameterSlice|DCUN1D1_T:USP30_T:ATP9B_N:DTD2_N:ERP44_N:SPN_N:STK39_N:affected:NoLN]
   [MeshVolume|MLF2_T:ATP8A1_N:FAM171A2_N:IAH1_N:PLEKHA3_N:RPL9_N:affected:NoLN]
   [SizeZoneNonUniformity|DCUN1D1_T:KRAS_T:PDAP1_T:PUM1_T:USP30_T:DTD2_N:PHLPP1_N:PLEKHA3_N:RPL9_N:STK39_N:affected:NoLN]
   [SmallDependenceHighGrayLevelEmphasis|DCUN1D1_T:KRAS_T:PDAP1_T:USP30_T:PHLPP1_N:affected:NoLN]
   [SurfaceArea|USP30_T:ERP44_N:FAM171A2_N:IAH1_N:SPN_N:affected:NoLN][TotalEnergy|COL21A1_N:PLEKHA3_N:RPL9_N:affected:NoLN]
   [VoxelVolume|MLF2_T:ATP8A1_N:FAM171A2_N:IAH1_N:PLEKHA3_N:RPL9_N:affected:NoLN]
  nodes:                                 40 
  arcs:                                  145 
    undirected arcs:                     0 
    directed arcs:                       145 
  average markov blanket size:           23.55 
  average neighbourhood size:            7.25 
  average branching factor:              3.62 

  learning algorithm:                    Hill-Climbing 
  score:                                 BIC (Gauss.) 
  penalization coefficient:              1.76318 
  tests used in the learning procedure:  2600 
  optimized:                             TRUE 

str.raw = boot.strength(imgFeatProtTable, R = 100, algorithm = "hc")#, algorithm.args = list(blacklist = bl, whitelist = wl))
attr(str.raw, "threshold")

[1] 0.5

avg.raw.full = averaged.network(str.raw)

Warning in averaged.network.backend(strength = strength, threshold = threshold) :
  arc TotalEnergy -> PLEKHA3_N would introduce cycles in the graph, ignoring.
Warning in averaged.network.backend(strength = strength, threshold = threshold) :
  arc COL21A1_N -> FAM171A2_N would introduce cycles in the graph, ignoring.
Warning in averaged.network.backend(strength = strength, threshold = threshold) :
  arc SurfaceArea -> LeastAxisLength would introduce cycles in the graph, ignoring.
Warning in averaged.network.backend(strength = strength, threshold = threshold) :
  arc COL21A1_N -> PLEKHA3_N would introduce cycles in the graph, ignoring.
Warning in averaged.network.backend(strength = strength, threshold = threshold) :
  arc SUMF2_T -> KRAS_T would introduce cycles in the graph, ignoring.
Warning in averaged.network.backend(strength = strength, threshold = threshold) :
  arc COL21A1_N -> PDE3A_N would introduce cycles in the graph, ignoring.
Warning in averaged.network.backend(strength = strength, threshold = threshold) :
  arc MeshVolume -> TotalEnergy would introduce cycles in the graph, ignoring.
Warning in averaged.network.backend(strength = strength, threshold = threshold) :
  arc MeshVolume -> VoxelVolume would introduce cycles in the graph, ignoring.

strength.plot(avg.raw.full, str.raw, shape = "ellipse")#, highlight = list(arcs = wl))

xval = bn.cv(imgFeatProtTable, bn = "hc", algorithm.args = list(blacklist = bl, whitelist = wl), loss = "cor-lw", loss.args = list(target = "affected", n = 200), runs = 10)

Warning in cor(data[, node], pred) : the standard deviation is zero
Warning in cor(data[, node], pred) : the standard deviation is zero
Warning in cor(data[, node], pred) : the standard deviation is zero
Warning in cor(data[, node], pred) : the standard deviation is zero
Warning in cor(data[, node], pred) : the standard deviation is zero
Warning in cor(data[, node], pred) : the standard deviation is zero
Warning in cor(data[, node], pred) : the standard deviation is zero
Warning in cor(data[, node], pred) : the standard deviation is zero
Warning in cor(data[, node], pred) : the standard deviation is zero
Warning in cor(data[, node], pred) : the standard deviation is zero
Warning in cor(data[, node], pred) : the standard deviation is zero
Warning in cor(data[, node], pred) : the standard deviation is zero
Warning in cor(data[, node], pred) : the standard deviation is zero
Warning in cor(data[, node], pred) : the standard deviation is zero
Warning in cor(data[, node], pred) : the standard deviation is zero
Warning in cor(data[, node], pred) : the standard deviation is zero
Warning in cor(data[, node], pred) : the standard deviation is zero
Warning in cor(data[, node], pred) : the standard deviation is zero
Warning in cor(data[, node], pred) : the standard deviation is zero

Compare different models and features

col_names <- Reduce(intersect, list(colnames(imgFeatTableF), colnames(tumProtTableF), colnames(normProtTableF), colnames(lnTableF)))

img_prot_af <- data.frame(t(data.frame(rbind(imgFeatTableF[, col_names], tumProtTableF[, col_names], normProtTableF[, col_names], lnAfTableF[, col_names]))))
img_tum_af <- data.frame(t(data.frame(rbind(imgFeatTableF[, col_names], tumProtTableF[, col_names], lnAfTableF[, col_names]))))
img_norm_af <- data.frame(t(data.frame(rbind(imgFeatTableF[, col_names], normProtTableF[, col_names], lnAfTableF[, col_names]))))
img_af <- data.frame(t(data.frame(rbind(imgFeatTableF[, col_names], lnAfTableF[, col_names]))))
prot_af <- data.frame(t(data.frame(rbind(tumProtTableF[, col_names], normProtTableF[, col_names], lnAfTableF[, col_names]))))

img_prot_no <- data.frame(t(data.frame(rbind(imgFeatTableF[, col_names], tumProtTableF[, col_names], normProtTableF[, col_names], lnNoTableF[, col_names]))))
img_no <- data.frame(t(data.frame(rbind(imgFeatTableF[, col_names], lnNoTableF[, col_names]))))
prot_no <- data.frame(t(data.frame(rbind(tumProtTableF[, col_names], normProtTableF[, col_names], lnNoTableF[, col_names]))))

bn_img_af <- empty.graph(c(rownames(imgFeatTableF), rownames(lnAfTableF)))
wl_img_af <- tiers2blacklist(list(rownames(lnAfTableF), rownames(imgFeatTableF)))
arcs(bn_img_af) <- wl_img_af

bn_prot_af <- empty.graph(c(rownames(tumProtTableF), rownames(normProtTableF), rownames(lnAfTableF)))
wl_prot_af <- tiers2blacklist(list(rownames(lnAfTableF), c(rownames(tumProtTableF), rownames(normProtTableF))))
arcs(bn_prot_af) <- wl_prot_af

bn_prot_img_af <- empty.graph(c(rownames(imgFeatTableF), rownames(tumProtTableF), rownames(normProtTableF), rownames(lnAfTableF)))
wl_prot_img_af <- rbind(
  tiers2blacklist(list(rownames(lnAfTableF), rownames(imgFeatTableF))),
  tiers2blacklist(list(rownames(imgFeatTableF), c(rownames(tumProtTableF), rownames(normProtTableF))))
)
arcs(bn_prot_img_af) <- wl_prot_img_af

bn_tum_img_af <- empty.graph(c(rownames(imgFeatTableF), rownames(tumProtTableF), rownames(lnAfTableF)))
wl_tum_img_af <- rbind(
  tiers2blacklist(list(rownames(lnAfTableF), rownames(imgFeatTableF))),
  tiers2blacklist(list(rownames(imgFeatTableF), c(rownames(tumProtTableF))))
)
arcs(bn_tum_img_af) <- wl_tum_img_af

bn_norm_img_af <- empty.graph(c(rownames(imgFeatTableF), rownames(normProtTableF), rownames(lnAfTableF)))
wl_norm_img_af <- rbind(
  tiers2blacklist(list(rownames(lnAfTableF), rownames(imgFeatTableF))),
  tiers2blacklist(list(rownames(imgFeatTableF), rownames(normProtTableF)))
)
arcs(bn_norm_img_af) <- wl_norm_img_af

bn_prot_img_id_af <- empty.graph(c(rownames(imgFeatTableF), rownames(tumProtTableF), rownames(normProtTableF), rownames(lnAfTableF)))
wl_prot_img_id_af <- tiers2blacklist(list(rownames(lnAfTableF), c(rownames(imgFeatTableF), rownames(tumProtTableF), rownames(normProtTableF))))
arcs(bn_prot_img_id_af) <- wl_prot_img_id_af

bn_img_no <- empty.graph(c(rownames(imgFeatTableF), rownames(lnNoTableF)))
wl_img_no <- tiers2blacklist(list(rownames(lnNoTableF), rownames(imgFeatTableF)))
arcs(bn_img_no) <- wl_img_no
bn_prot_no <- empty.graph(c(rownames(tumProtTableF), rownames(normProtTableF), rownames(lnNoTableF)))
wl_prot_no <- tiers2blacklist(list(rownames(lnNoTableF), c(rownames(tumProtTableF), rownames(normProtTableF))))
arcs(bn_prot_no) <- wl_prot_no
bn_prot_img_no <- empty.graph(c(rownames(imgFeatTableF), rownames(tumProtTableF), rownames(normProtTableF), rownames(lnNoTableF)))
wl_prot_img_no <- rbind(
  tiers2blacklist(list(rownames(lnNoTableF), rownames(imgFeatTableF))),
  tiers2blacklist(list(rownames(imgFeatTableF), c(rownames(tumProtTableF), rownames(normProtTableF))))
)
arcs(bn_prot_img_no) <- wl_prot_img_no
bn_prot_img_id_no <- empty.graph(c(rownames(imgFeatTableF), rownames(tumProtTableF), rownames(normProtTableF), rownames(lnNoTableF)))
wl_prot_img_id_no <- tiers2blacklist(list(rownames(lnNoTableF), c(rownames(imgFeatTableF), rownames(tumProtTableF), rownames(normProtTableF))))
arcs(bn_prot_img_id_no) <- wl_prot_img_id_no

graphviz.plot(bn_img_af)

graphviz.plot(bn_prot_af)

graphviz.plot(bn_prot_img_af)

graphviz.plot(bn_prot_img_id_af)

graphviz.plot(bn_tum_img_af)

graphviz.plot(bn_norm_img_af)

img_af_cv <- bn.cv(img_af, bn = bn_img_af, runs = 10, loss = "mse-lw", loss.args = list(target = "affected"))
prot_af_cv <- bn.cv(prot_af, bn = bn_prot_af, runs = 10, loss = "mse-lw", loss.args = list(target = "affected"))
img_prot_af_cv <- bn.cv(img_prot_af, bn = bn_prot_img_af, runs = 10, loss = "mse-lw", loss.args = list(target = "affected"))
img_tum_af_cv <- bn.cv(img_tum_af, bn = bn_tum_img_af, runs = 10, loss = "mse-lw", loss.args = list(target = "affected"))
img_norm_af_cv <- bn.cv(img_norm_af, bn = bn_norm_img_af, runs = 10, loss = "mse-lw", loss.args = list(target = "affected"))
img_prot_id_af_cv <- bn.cv(img_prot_af, bn = bn_prot_img_id_af, runs = 10, loss = "mse-lw", loss.args = list(target = "affected"))

plot(img_af_cv, prot_af_cv, img_prot_af_cv, img_tum_af_cv, img_norm_af_cv, img_prot_id_af_cv, xlab = c("Image only", "Proteomics only", "Proteomics -> Image", "Tum prot -> Image", "Norm prot -> Image", "Image + Proteomics"))

plot(img_af_cv, img_prot_af_cv, xlab = c("Image only", "Proteomics -> Image"))

plot(img_af_cv, img_prot_af_cv, img_tum_af_cv, img_norm_af_cv, xlab = c("Image only", "Proteomics -> Image", "Tum prot -> Image", "Norm prot -> Image"))

save(img_af_cv, prot_af_cv, img_prot_af_cv, img_tum_af_cv, img_norm_af_cv, img_prot_id_af_cv, file = "./analysis_SF/comparisons.RData")

valid learning algorithm(s) are “gs” (Grow-Shrink), “iamb” (IAMB), “fast.iamb” (Fast-IAMB), “inter.iamb” (Inter-IAMB), “iamb.fdr” (IAMB-FDR), “pc.stable” (PC (Stable)), “mmpc” (Max-Min Parent Children), “si.hiton.pc” (Semi-Interleaved HITON-PC), “hpc” (Hybrid Parents and Children), “hc” (Hill-Climbing), “tabu” (Tabu Search), “rsmax2” (Two-Phase Restricted Maximization), “mmhc” (Max-Min Hill-Climbing), “h2pc” (Hybrid^2 Parent Children), “chow.liu” (Chow-Liu), “aracne” (ARACNE), “naive.bayes” (Naive Bayes Classifier), “tree.bayes” (TAN Bayes Classifier). See ?bnlearn-package for details.

img_af_ho <- bn.cv(img_af, bn = bn_img_af, runs = 10, method = "hold-out", k = 10, m = 10, loss = "mse-lw", loss.args = list(target = "affected"))
prot_af_ho <- bn.cv(prot_af, bn = bn_prot_af, runs = 10, method = "hold-out", k = 10, m = 10, loss = "mse-lw", loss.args = list(target = "affected"))
img_prot_af_ho <- bn.cv(img_prot_af, bn = bn_prot_img_af, runs = 10, method = "hold-out", k = 10, m = 10, loss = "mse-lw", loss.args = list(target = "affected"))
img_tum_af_ho <- bn.cv(img_tum_af, bn = bn_tum_img_af, runs = 10, method = "hold-out", k = 10, m = 10, loss = "mse-lw", loss.args = list(target = "affected"))
img_norm_af_ho <- bn.cv(img_norm_af, bn = bn_norm_img_af, runs = 10, method = "hold-out", k = 10, m = 10, loss = "mse-lw", loss.args = list(target = "affected"))
img_prot_id_af_ho <- bn.cv(img_prot_af, bn = bn_prot_img_id_af, runs = 10, method = "hold-out", k = 10, m = 10, loss = "mse-lw", loss.args = list(target = "affected"))
plot(img_af_ho, prot_af_ho, img_prot_af_ho, img_tum_af_ho, img_norm_af_ho, img_prot_id_af_ho, xlab = c("Image only", "Proteomics only", "Proteomics -> Image", "Tum prot -> Image", "Norm prot -> Image", "Image + Proteomics"))
plot(img_af_ho, img_prot_af_ho, xlab = c("Image only", "Proteomics -> Image"))
plot(img_af_ho, img_prot_af_ho, img_tum_af_ho, img_norm_af_ho, xlab = c("Image only", "Proteomics -> Image", "Tum prot -> Image", "Norm prot -> Image"))

img_af_cv <- bn.cv(img_af, bn = bn_img_af, runs = 10, loss = "mse-lw", loss.args = list(target = "affected"))
prot_af_cv <- bn.cv(prot_af, bn = bn_prot_af, runs = 10, loss = "mse-lw", loss.args = list(target = "affected"))
img_prot_af_cv <- bn.cv(img_prot_af, bn = bn_prot_img_af, runs = 10, loss = "mse-lw", loss.args = list(target = "affected"))
img_prot_id_af_cv <- bn.cv(img_prot_af, bn = bn_prot_img_id_af, runs = 10, loss = "mse-lw", loss.args = list(target = "affected"))
plot(img_af_cv, prot_af_cv, img_prot_af_cv, img_prot_id_af_cv, xlab = c("Image only", "Proteomics only", "Proteomics -> Image", "Image + Proteomics"))
plot(img_af_cv, img_prot_af_cv, xlab = c("Image only", "Proteomics -> Image"))

img_af_cv <- bn.cv(img_af, bn = bn_img_af, runs = 10, loss = "mse", loss.args = list(target = "affected"))
prot_af_cv <- bn.cv(prot_af, bn = bn_prot_af, runs = 10, loss = "mse", loss.args = list(target = "affected"))
img_prot_af_cv <- bn.cv(img_prot_af, bn = bn_prot_img_af, runs = 10, loss = "mse", loss.args = list(target = "affected"))
img_prot_id_af_cv <- bn.cv(img_prot_af, bn = bn_prot_img_id_af, runs = 10, loss = "mse", loss.args = list(target = "affected"))
plot(img_af_cv, prot_af_cv, img_prot_af_cv, img_prot_id_af_cv, xlab = c("Image only", "Proteomics only", "Proteomics -> Image", "Image + Proteomics"))
plot(img_af_cv, img_prot_af_cv, xlab = c("Image only", "Proteomics -> Image"))

img_no_cv <- bn.cv(img_no, bn = bn_img_no, runs = 10, loss = "mse", loss.args = list(target = "NoLN"))
prot_no_cv <- bn.cv(prot_no, bn = bn_prot_no, runs = 10, loss = "mse", loss.args = list(target = "NoLN"))
img_prot_no_cv <- bn.cv(img_prot_no, bn = bn_prot_img_no, runs = 10, loss = "mse", loss.args = list(target = "NoLN"))
img_prot_id_no_cv <- bn.cv(img_prot_no, bn = bn_prot_img_id_no, runs = 10, loss = "mse", loss.args = list(target = "NoLN"))
plot(img_no_cv, prot_no_cv, img_prot_no_cv, img_prot_id_no_cv, xlab = c("Image only", "Proteomics only", "Proteomics -> Image", "Image + Proteomics"))
plot(img_no_cv, img_prot_no_cv, xlab = c("Image only", "Proteomics -> Image"))

Tests

img_af_cv <- bn.cv(img_af, bn = "hc", runs = 10, loss = "mse", loss.args = list(target = "affected"))
prot_af_cv <- bn.cv(prot_af, bn = "hc", runs = 10, loss = "mse", loss.args = list(target = "affected"))
img_prot_af_cv <- bn.cv(img_prot_af, bn = "hc", runs = 10, loss = "mse", loss.args = list(target = "affected"))
plot(img_af_cv, prot_af_cv, img_prot_af_cv, xlab = c("Image only", "Proteomics only", "Image + Proteomics"))

img_no_cv <- bn.cv(img_no, bn = "tabu", runs = 10, loss = "mse", loss.args = list(target = "NoLN"))
prot_no_cv <- bn.cv(prot_no, bn = "tabu", runs = 10, loss = "mse", loss.args = list(target = "NoLN"))
img_prot_no_cv <- bn.cv(img_prot_no, bn = "tabu", runs = 10, loss = "mse", loss.args = list(target = "NoLN"))
plot(img_no_cv, prot_no_cv, img_prot_no_cv, xlab = c("Image only", "Proteomics only", "Image + Proteomics"))

img_no_cv <- bn.cv(img_no, bn = "hc", runs = 10, loss = "mse", loss.args = list(target = "NoLN"))
prot_no_cv <- bn.cv(prot_no, bn = "hc", runs = 10, loss = "mse", loss.args = list(target = "NoLN"))
img_prot_no_cv <- bn.cv(img_prot_no, bn = "hc", runs = 10, loss = "mse", loss.args = list(target = "NoLN"))
plot(img_no_cv, prot_no_cv, img_prot_no_cv, xlab = c("Image only", "Proteomics only", "Image + Proteomics"))

Using other models to do the cross validation, like AutoML?

Hidden Markov Model for modeling the lymphinode dynamics